minmax-ravel-benchmark.md

Benchmark: Tensor.minmax() vs np.min + np.max on image-shaped data

Context: Evaluation for image augmentation / vision pipelines (e.g. Albucore). We compared NumKong Tensor.minmax() to two-pass NumPy (min + max) on realistic (H, W, C) layouts (channel-last), C-contiguous, flattened to 1D — i.e. global min and max over all pixels, not per-channel axis reductions.

Takeaway (below hardware): On this setup, NumPy was faster for every shape/dtype tested. We are not planning to route global min+max through NumKong minmax() for contiguous flat buffers unless future versions or other CPUs change that.

---

Reference snippet (Python)

import time

import numkong as nk
import numpy as np


def median_ms(fn, repeats=11, warmup=3):
    for _ in range(warmup):
        fn()
    times = []
    for _ in range(repeats):
        t0 = time.perf_counter()
        fn()
        times.append((time.perf_counter() - t0) * 1000)
    return float(np.median(times))


rng = np.random.default_rng(0)
for dtype, name in [(np.float32, "float32"), (np.uint8, "uint8")]:
    for h, w, c in [(224, 224, 3), (512, 512, 3), (1024, 1024, 1)]:
        if dtype == np.uint8:
            img = rng.integers(0, 256, size=(h, w, c), dtype=np.uint8)
        else:
            img = rng.random((h, w, c), dtype=np.float32)
        flat = np.ascontiguousarray(img).reshape(-1)
        t = nk.Tensor(flat)
        t_nk = median_ms(lambda: t.minmax())
        t_np = median_ms(lambda: (flat.min(), flat.max()))
        ratio = max(t_nk, t_np) / min(t_nk, t_np)
        winner = "NumPy" if t_np <= t_nk else "NumKong"
        print(f"{name}  {h}x{w}x{c}  n={flat.size}  nk={t_nk:.4f}ms  np={t_np:.4f}ms  -> {winner}  {ratio:.2f}x")

---

Results (median ms, 11 repeats after 3 warmup, seed=0)

dtype	shape (H×W×C)	pixels	NumKong minmax (ms)	np.min + np.max (ms)	faster	ratio
float32	224 × 224 × 3	150528	0.0336	0.0112	NumPy	3.00×
float32	512 × 512 × 3	786432	0.1746	0.0543	NumPy	3.22×
float32	1024 × 1024 × 1	1048576	0.2328	0.0728	NumPy	3.20×
uint8	224 × 224 × 3	150528	0.0048	0.0038	NumPy	1.24×
uint8	512 × 512 × 3	786432	0.0270	0.0145	NumPy	1.86×
uint8	1024 × 1024 × 1	1048576	0.0325	0.0210	NumPy	1.55×

---

Environment

• OS: Darwin, arm64
• CPU: Apple M4 Max
• numkong 7.0.0
• numpy 2.4.2

Caveat: This does not benchmark axis-wise minmax (e.g. per column / per channel along spatial axes). NumKong may still win on other access patterns (e.g. strided column reductions); this table is only global min+max on a contiguous 1D view of an image tensor.